Fig. 6

Key steps for unsupervised exposure of the internal structure of neuronal activity and the reconstruction of internal representations. First, the population neuronal activity is binarized and temporally binned into N-dimensional activity vectors (N = number of neurons). A dimensionality reduction algorithm is then applied to the activity vectors, embedding the data and exposing its internal structure within a low-dimensional space. Next, the activity in the reduced space is parameterized using a clustering algorithm. This allows the reconstruction of the trajectory of neuronal activity and the internal tuning curves of individual neurons as a function of the parameterized variable. In parallel, the dimension and topology of the data in the reduced space can be estimated to reveal key properties of the internal representations